stGCL: A versatile cross-modality fusion method based on multi-modal graph contrastive learning for spatial transcriptomics

Abstract

Advances in spatial transcriptomics (ST) technologies have provided unprecedented opportunities to depict transcriptomic and histological landscapes in the spatial context. Multi-modal ST data provide abundant and comprehensive information about cellular status, function, and organization. However, in dealing with the processing and analysis of spatial transcriptomics data, existing algorithms struggle to effectively fuse the multi-modal information contained within ST data. Here, we propose a graph contrastive learning-based cross-modality fusion model named stGCL for accurate and robust integrating gene expression, spatial information as well as histological profiles simultaneously. stGCL adopts a novel histology-based Vision Transformer (H-ViT) method to effectively encode histological features and combines multi-modal graph attention auto-encoder (GATE) with contrastive learning to fuse cross-modality features. In addition, stGCL introduces a pioneering spatial coordinate correcting and registering strategy for tissue slices integration, which can reduce batch effects and identify cross-sectional domains precisely. Compared with state-of-the-art methods on spatial transcriptomics data across platforms and resolutions, stGCL achieves a superior clustering performance and is more robust in unraveling spatial patterns of biological significance. Additionally, stGCL successfully reconstructed three-dimensional (3D) brain tissue structures by integrating vertical and horizontal slices respectively. Application of stGCL in human bronchiolar adenoma (BA) data reveals intratumor spatial heterogeneity and identifies candidate gene biomarkers. In summary, stGCL enables the fusion of various spatial modality data and is a powerful tool for analytical tasks such as spatial domain identification and multi-slice integration.

STEP: Spatial Transcriptomics Embedding Procedure for Multi-scale Biological Heterogeneities Revelation in Multiple Samples

“…Cell clustering, aiming to differentiate cells by cell types (Chidester et al, 2023; Li et al, 2022; Miller et al, 2021; Teng et al, 2022); 2. Spatial domain identification for discovery of biologically functional regions with a certain degree of spatial continuity in tissues (Dong & Zhang, 2022; Hu et al, 2021; Hu et al, 2024; Li & Zhou, 2022; Liu et al, 2023; Long et al, 2023; Shang & Zhou, 2022; Yang et al, 2024; Yu et al, 2023; Zhang et al, 2024; Zhao et al, 2021). The significant improvements can be observed on solving these two tasks in the single section scenario for a specific data resolution.…”

“…A batch-embedding based paired transformation model (BEM), introducing an unparalleled batch-correction technique that operates independently from BBM and delineates a new direction in handling batch variations; and 3. A spatial model (SpM) leveraging the graph convolutional network (GCN) (Kipf & Welling, 2017) distinctively applied to low-dimensional embeddings as a spatial filter or smoother rather than direct gene expression data as a feature extractor (Dong & Zhang, 2022; Gao et al, 2024; Hu et al, 2021; Long et al, 2023; Yu et al, 2023; Zhang et al, 2024; Zhou et al, 2023)—a first in the field. This innovative use of GCN significantly reduces noise, effectively capturing spatial contexts and variations with enhanced biological relevance.…”

“…Not only are these assumptions incompletely correct, they are actually unnecessary because cell types can be well distin-guished at the level of gene expression regardless of spatial locations. On the other hand, when it comes to the integration of multiple sections, current methods often require the spatial contiguity between sections, whether vertically or horizontally (Clifton et al, 2023; Dong & Zhang, 2022; Gao et al, 2024; Long et al, 2023; Wang et al, 2023; Yu et al, 2023; Yuan, 2024; Zhou et al, 2023). Though, in concept, spatial domains demon-strate the spatial variations and the lower resolution compared to the cell types, the most fundamental variations that segment the regions are still transcriptional.…”

“…Not only are these assumptions incompletely correct, they are actually unnecessary because cell types can be well distinguished at the level of gene expression regardless of spatial locations. On the other hand, when it comes to the integration of multiple sections, current methods often require the spatial contiguity between sections, whether vertically or horizontally (Clifton et al, 2023;Dong & Zhang, 2022;Gao et al, 2024;Long et al, 2023;Wang et al, 2023;Yu et al, 2023;Yuan, 2024;Zhou et al, 2023).Though, in concept, spatial domains demonstrate the spatial variations and the lower resolution compared to the cell types, the most fundamental variations that segment the regions are still transcriptional. Although one Bayesian model (Li & Zhou, 2022) and a cell-type annotation based method (Yuan, 2024) can handle both the multi-scale or multi-section (non-contiguous) SRT dataset, there is another limitation unique to these methods: they are unable to provide embeddings that are consistent to their identified spatial domains for the downstream analyses, like pseudotime analysis, where the PCA and other batch-effect removal algorithms specific for scRNA-seq data (Haghverdi et al, 2018;Korsunsky et al, 2019) are conducted, instead.…”